| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 91 === Generally, eigenvoice based speaker adaptation is effective when the adaptation data is insufficient. This thesis presents a novel approach to eigenvoice speaker adaptation and applies the proposed approach to build a verbal information verification (VIV) system of Mandarin Chinese.
First of all, we present the maximum a posteriori eigen- decomposition (MAPED), where the linear combination coefficients for eigenvector decomposition are estimated according MAP estimation. By considering the prior density of combination coefficient as a Gaussian distribution, the MAP estimate of coefficient is obtained to carry out MAPED. In case of insufficient data, MAPED is able to achieve better estimate than maximum likelihood eigen-decomposition (MLED).
On the other hand, we use the principal component analysis (PCA) to project the speaker-specific hidden Markov model (HMM) parameters of each state onto a smaller orthogonal feature space. Then, we calculate the HMM covariance matrix using the observations in new feature space. The adaptive HMM covariance matrix is estimated by transforming matrix to the original feature space.
From the experiments, we find that the methods of adaptive HMM covariance matrix and MAPED for eigenvoice speaker adaptation can significantly improve Mandarin speech recognition. Furthermore, we apply the proposed eigenvoice adaptation method to build a verbal information verification (VIV) demonstration system. The system performance is greatly improved by incorporating the speaker enrollment phase where the HMM parameters of new speaker are well adjusted.
|
|---|
